def prep_bbox(sess,
x,
y,
x_train,
y_train,
x_test,
y_test,
nb_epochs,
batch_size,
learning_rate,
rng,
nb_classes=10,
img_rows=28,
img_cols=28,
nchannels=1):
"""
Define and train a model that simulates the "remote"
black-box oracle described in the original paper.
:param sess: the TF session
:param x: the input placeholder for MNIST
:param y: the ouput placeholder for MNIST
:param x_train: the training data for the oracle
:param y_train: the training labels for the oracle
:param x_test: the testing data for the oracle
:param y_test: the testing labels for the oracle
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param learning_rate: learning rate for training
:param rng: numpy.random.RandomState
:return:
"""
# Define TF model graph (for the black-box model)
nb_filters = 64
model = ModelBasicCNN('model1', nb_classes, nb_filters)
loss = CrossEntropy(model, smoothing=0.1)
predictions = model.get_logits(x)
print("Defined TensorFlow model graph.")
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
train(sess, loss, x, y, x_train, y_train, args=train_params, rng=rng)
# Print out the accuracy on legitimate data
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess,
x,
y,
predictions,
x_test,
y_test,
args=eval_params)
print('Test accuracy of black-box on legitimate test '
'examples: ' + str(accuracy))
return model, predictions, accuracy
def train_sub(sess,
x,
y,
bbox_preds,
x_sub,
y_sub,
nb_classes,
nb_epochs_s,
batch_size,
learning_rate,
data_aug,
lmbda,
aug_batch_size,
rng,
img_rows=28,
img_cols=28,
nchannels=1):
"""
This function creates the substitute by alternatively
augmenting the training data and training the substitute.
:param sess: TF session
:param x: input TF placeholder
:param y: output TF placeholder
:param bbox_preds: output of black-box model predictions
:param x_sub: initial substitute training data
:param y_sub: initial substitute training labels
:param nb_classes: number of output classes
:param nb_epochs_s: number of epochs to train substitute model
:param batch_size: size of training batches
:param learning_rate: learning rate for training
:param data_aug: number of times substitute training data is augmented
:param lmbda: lambda from arxiv.org/abs/1602.02697
:param rng: numpy.random.RandomState instance
:return:
"""
# Define TF model graph (for the black-box model)
model_sub = ModelSubstitute('model_s', nb_classes)
preds_sub = model_sub.get_logits(x)
loss_sub = CrossEntropy(model_sub, smoothing=0)
print("Defined TensorFlow model graph for the substitute.")
# Define the Jacobian symbolically using TensorFlow
grads = jacobian_graph(preds_sub, x, nb_classes)
# Train the substitute and augment dataset alternatively
for rho in xrange(data_aug):
print("Substitute training epoch #" + str(rho))
train_params = {
'nb_epochs': nb_epochs_s,
'batch_size': batch_size,
'learning_rate': learning_rate
}
with TemporaryLogLevel(logging.WARNING, "cleverhans.utils.tf"):
train(sess,
loss_sub,
x,
y,
x_sub,
to_categorical(y_sub, nb_classes),
init_all=False,
args=train_params,
rng=rng,
var_list=model_sub.get_params())
# If we are not at last substitute training iteration, augment dataset
if rho < data_aug - 1:
print("Augmenting substitute training data.")
# Perform the Jacobian augmentation
lmbda_coef = 2 * int(int(rho / 3) != 0) - 1
x_sub = jacobian_augmentation(sess, x, x_sub, y_sub, grads,
lmbda_coef * lmbda, aug_batch_size)
print("Labeling substitute training data.")
# Label the newly generated synthetic points using the black-box
y_sub = np.hstack([y_sub, y_sub])
x_sub_prev = x_sub[int(len(x_sub) / 2):]
eval_params = {'batch_size': batch_size}
bbox_val = batch_eval(sess, [x], [bbox_preds], [x_sub_prev],
args=eval_params)[0]
# Note here that we take the argmax because the adversary
# only has access to the label (not the probabilities) output
# by the black-box model
y_sub[int(len(x_sub) / 2):] = np.argmax(bbox_val, axis=1)
return model_sub, preds_sub
def mnist_tutorial(train_start=0, train_end=60000, test_start=0,
test_end=10000, nb_epochs=NB_EPOCHS, batch_size=BATCH_SIZE,
learning_rate=LEARNING_RATE,
clean_train=CLEAN_TRAIN,
testing=False,
backprop_through_attack=BACKPROP_THROUGH_ATTACK,
nb_filters=NB_FILTERS, num_threads=None,
label_smoothing=0.1):
"""
MNIST cleverhans tutorial
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param learning_rate: learning rate for training
:param clean_train: perform normal training on clean examples only
before performing adversarial training.
:param testing: if true, complete an AccuracyReport for unit tests
to verify that performance is adequate
:param backprop_through_attack: If True, backprop through adversarial
example construction process during
adversarial training.
:param label_smoothing: float, amount of label smoothing for cross entropy
:return: an AccuracyReport object
"""
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Set logging level to see debug information
set_log_level(logging.DEBUG)
# Create TF session
if num_threads:
config_args = dict(intra_op_parallelism_threads=1)
else:
config_args = {}
sess = tf.Session(config=tf.ConfigProto(**config_args))
# Get MNIST test data
x_train, y_train, x_test, y_test = data_mnist(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
# Use Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols,
nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
eval_params = {'batch_size': batch_size}
fgsm_params = {
'eps': 0.3,
'clip_min': 0.,
'clip_max': 1.
}
rng = np.random.RandomState([2017, 8, 30])
def do_eval(preds, x_set, y_set, report_key, is_adv=None):
acc = model_eval(sess, x, y, preds, x_set, y_set, args=eval_params)
setattr(report, report_key, acc)
if is_adv is None:
report_text = None
elif is_adv:
report_text = 'adversarial'
else:
report_text = 'legitimate'
if report_text:
print('Test accuracy on %s examples: %0.4f' % (report_text, acc))
if clean_train:
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=label_smoothing)
def evaluate():
do_eval(preds, x_test, y_test, 'clean_train_clean_eval', False)
train(sess, loss, x, y, x_train, y_train, evaluate=evaluate,
args=train_params, rng=rng, var_list=model.get_params())
# Calculate training error
if testing:
do_eval(preds, x_train, y_train, 'train_clean_train_clean_eval')
# Initialize the Fast Gradient Sign Method (FGSM) attack object and
# graph
fgsm = FastGradientMethod(model, sess=sess)
adv_x = fgsm.generate(x, **fgsm_params)
preds_adv = model.get_logits(adv_x)
# Evaluate the accuracy of the MNIST model on adversarial examples
do_eval(preds_adv, x_test, y_test, 'clean_train_adv_eval', True)
# Calculate training error
if testing:
do_eval(preds_adv, x_train, y_train, 'train_clean_train_adv_eval')
print('Repeating the process, using adversarial training')
# Create a new model and train it to be robust to FastGradientMethod
model2 = ModelBasicCNN('model2', nb_classes, nb_filters)
fgsm2 = FastGradientMethod(model2, sess=sess)
def attack(x):
return fgsm2.generate(x, **fgsm_params)
loss2 = CrossEntropy(model2, smoothing=label_smoothing, attack=attack)
preds2 = model2.get_logits(x)
adv_x2 = attack(x)
if not backprop_through_attack:
# For the fgsm attack used in this tutorial, the attack has zero
# gradient so enabling this flag does not change the gradient.
# For some other attacks, enabling this flag increases the cost of
# training, but gives the defender the ability to anticipate how
# the atacker will change their strategy in response to updates to
# the defender's parameters.
adv_x2 = tf.stop_gradient(adv_x2)
preds2_adv = model2.get_logits(adv_x2)
def evaluate2():
# Accuracy of adversarially trained model on legitimate test inputs
do_eval(preds2, x_test, y_test, 'adv_train_clean_eval', False)
# Accuracy of the adversarially trained model on adversarial examples
do_eval(preds2_adv, x_test, y_test, 'adv_train_adv_eval', True)
# Perform and evaluate adversarial training
train(sess, loss2, x, y, x_train, y_train, evaluate=evaluate2,
args=train_params, rng=rng, var_list=model2.get_params())
# Calculate training errors
if testing:
do_eval(preds2, x_train, y_train, 'train_adv_train_clean_eval')
do_eval(preds2_adv, x_train, y_train, 'train_adv_train_adv_eval')
return report
def mnist_tutorial_jsma(train_start=0,
train_end=60000,
test_start=0,
test_end=10000,
viz_enabled=VIZ_ENABLED,
nb_epochs=NB_EPOCHS,
batch_size=BATCH_SIZE,
source_samples=SOURCE_SAMPLES,
learning_rate=LEARNING_RATE):
"""
MNIST tutorial for the Jacobian-based saliency map approach (JSMA)
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param viz_enabled: (boolean) activate plots of adversarial examples
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param nb_classes: number of output classes
:param source_samples: number of test inputs to attack
:param learning_rate: learning rate for training
:return: an AccuracyReport object
"""
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Create TF session and set as Keras backend session
sess = tf.Session()
print("Created TensorFlow session.")
set_log_level(logging.DEBUG)
# Get MNIST test data
x_train, y_train, x_test, y_test = data_mnist(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
# Obtain Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols, nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
nb_filters = 64
# Define TF model graph
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=0.1)
print("Defined TensorFlow model graph.")
###########################################################################
# Training the model using TensorFlow
###########################################################################
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate
}
sess.run(tf.global_variables_initializer())
rng = np.random.RandomState([2017, 8, 30])
train(sess, loss, x, y, x_train, y_train, args=train_params, rng=rng)
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
assert x_test.shape[0] == test_end - test_start, x_test.shape
print('Test accuracy on legitimate test examples: {0}'.format(accuracy))
report.clean_train_clean_eval = accuracy
###########################################################################
# Craft adversarial examples using the Jacobian-based saliency map approach
###########################################################################
print('Crafting ' + str(source_samples) + ' * ' + str(nb_classes - 1) +
' adversarial examples')
# Keep track of success (adversarial example classified in target)
results = np.zeros((nb_classes, source_samples), dtype='i')
# Rate of perturbed features for each test set example and target class
perturbations = np.zeros((nb_classes, source_samples), dtype='f')
# Initialize our array for grid visualization
grid_shape = (nb_classes, nb_classes, img_rows, img_cols, nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
# Instantiate a SaliencyMapMethod attack object
jsma = SaliencyMapMethod(model, back='tf', sess=sess)
jsma_params = {
'theta': 1.,
'gamma': 0.1,
'clip_min': 0.,
'clip_max': 1.,
'y_target': None
}
figure = None
# Loop over the samples we want to perturb into adversarial examples
for sample_ind in xrange(0, source_samples):
print('--------------------------------------')
print('Attacking input %i/%i' % (sample_ind + 1, source_samples))
sample = x_test[sample_ind:(sample_ind + 1)]
# We want to find an adversarial example for each possible target class
# (i.e. all classes that differ from the label given in the dataset)
current_class = int(np.argmax(y_test[sample_ind]))
target_classes = other_classes(nb_classes, current_class)
# For the grid visualization, keep original images along the diagonal
grid_viz_data[current_class, current_class, :, :, :] = np.reshape(
sample, (img_rows, img_cols, nchannels))
# Loop over all target classes
for target in target_classes:
print('Generating adv. example for target class %i' % target)
# This call runs the Jacobian-based saliency map approach
one_hot_target = np.zeros((1, nb_classes), dtype=np.float32)
one_hot_target[0, target] = 1
jsma_params['y_target'] = one_hot_target
adv_x = jsma.generate_np(sample, **jsma_params)
# Check if success was achieved
res = int(model_argmax(sess, x, preds, adv_x) == target)
# Computer number of modified features
adv_x_reshape = adv_x.reshape(-1)
test_in_reshape = x_test[sample_ind].reshape(-1)
nb_changed = np.where(adv_x_reshape != test_in_reshape)[0].shape[0]
percent_perturb = float(nb_changed) / adv_x.reshape(-1).shape[0]
# Display the original and adversarial images side-by-side
if viz_enabled:
figure = pair_visual(
np.reshape(sample, (img_rows, img_cols, nchannels)),
np.reshape(adv_x, (img_rows, img_cols, nchannels)), figure)
# Add our adversarial example to our grid data
grid_viz_data[target, current_class, :, :, :] = np.reshape(
adv_x, (img_rows, img_cols, nchannels))
# Update the arrays for later analysis
results[target, sample_ind] = res
perturbations[target, sample_ind] = percent_perturb
print('--------------------------------------')
# Compute the number of adversarial examples that were successfully found
nb_targets_tried = ((nb_classes - 1) * source_samples)
succ_rate = float(np.sum(results)) / nb_targets_tried
print('Avg. rate of successful adv. examples {0:.4f}'.format(succ_rate))
report.clean_train_adv_eval = 1. - succ_rate
# Compute the average distortion introduced by the algorithm
percent_perturbed = np.mean(perturbations)
print('Avg. rate of perturbed features {0:.4f}'.format(percent_perturbed))
# Compute the average distortion introduced for successful samples only
percent_perturb_succ = np.mean(perturbations * (results == 1))
print('Avg. rate of perturbed features for successful '
'adversarial examples {0:.4f}'.format(percent_perturb_succ))
# Close TF session
sess.close()
# Finally, block & display a grid of all the adversarial examples
if viz_enabled:
import matplotlib.pyplot as plt
plt.close(figure)
_ = grid_visual(grid_viz_data)
return report
def mnist_tutorial_cw(train_start=0, train_end=60000, test_start=0,
test_end=10000, viz_enabled=VIZ_ENABLED,
nb_epochs=NB_EPOCHS, batch_size=BATCH_SIZE,
source_samples=SOURCE_SAMPLES,
learning_rate=LEARNING_RATE,
attack_iterations=ATTACK_ITERATIONS,
model_path=MODEL_PATH,
targeted=TARGETED):
"""
MNIST tutorial for Carlini and Wagner's attack
:param train_start: index of first training set example
:param train_end: index of last training set example
:param test_start: index of first test set example
:param test_end: index of last test set example
:param viz_enabled: (boolean) activate plots of adversarial examples
:param nb_epochs: number of epochs to train model
:param batch_size: size of training batches
:param nb_classes: number of output classes
:param source_samples: number of test inputs to attack
:param learning_rate: learning rate for training
:param model_path: path to the model file
:param targeted: should we run a targeted attack? or untargeted?
:return: an AccuracyReport object
"""
# Object used to keep track of (and return) key accuracies
report = AccuracyReport()
# Set TF random seed to improve reproducibility
tf.set_random_seed(1234)
# Create TF session
sess = tf.Session()
print("Created TensorFlow session.")
set_log_level(logging.DEBUG)
# Get MNIST test data
x_train, y_train, x_test, y_test = data_mnist(train_start=train_start,
train_end=train_end,
test_start=test_start,
test_end=test_end)
# Obtain Image Parameters
img_rows, img_cols, nchannels = x_train.shape[1:4]
nb_classes = y_train.shape[1]
# Define input TF placeholder
x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols,
nchannels))
y = tf.placeholder(tf.float32, shape=(None, nb_classes))
nb_filters = 64
# Define TF model graph
model = ModelBasicCNN('model1', nb_classes, nb_filters)
preds = model.get_logits(x)
loss = CrossEntropy(model, smoothing=0.1)
print("Defined TensorFlow model graph.")
###########################################################################
# Training the model using TensorFlow
###########################################################################
# Train an MNIST model
train_params = {
'nb_epochs': nb_epochs,
'batch_size': batch_size,
'learning_rate': learning_rate,
'filename': os.path.split(model_path)[-1]
}
rng = np.random.RandomState([2017, 8, 30])
# check if we've trained before, and if we have, use that pre-trained model
if os.path.exists(model_path + ".meta"):
tf_model_load(sess, model_path)
else:
train(sess, loss, x, y, x_train, y_train, args=train_params, rng=rng)
saver = tf.train.Saver()
saver.save(sess, model_path)
# Evaluate the accuracy of the MNIST model on legitimate test examples
eval_params = {'batch_size': batch_size}
accuracy = model_eval(sess, x, y, preds, x_test, y_test, args=eval_params)
assert x_test.shape[0] == test_end - test_start, x_test.shape
print('Test accuracy on legitimate test examples: {0}'.format(accuracy))
report.clean_train_clean_eval = accuracy
###########################################################################
# Craft adversarial examples using Carlini and Wagner's approach
###########################################################################
nb_adv_per_sample = str(nb_classes - 1) if targeted else '1'
print('Crafting ' + str(source_samples) + ' * ' + nb_adv_per_sample +
' adversarial examples')
print("This could take some time ...")
# Instantiate a CW attack object
cw = CarliniWagnerL2(model, back='tf', sess=sess)
if viz_enabled:
assert source_samples == nb_classes
idxs = [np.where(np.argmax(y_test, axis=1) == i)[0][0]
for i in range(nb_classes)]
if targeted:
if viz_enabled:
# Initialize our array for grid visualization
grid_shape = (nb_classes, nb_classes, img_rows, img_cols,
nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
adv_inputs = np.array(
[[instance] * nb_classes for instance in x_test[idxs]],
dtype=np.float32)
else:
adv_inputs = np.array(
[[instance] * nb_classes for
instance in x_test[:source_samples]], dtype=np.float32)
one_hot = np.zeros((nb_classes, nb_classes))
one_hot[np.arange(nb_classes), np.arange(nb_classes)] = 1
adv_inputs = adv_inputs.reshape(
(source_samples * nb_classes, img_rows, img_cols, nchannels))
adv_ys = np.array([one_hot] * source_samples,
dtype=np.float32).reshape((source_samples *
nb_classes, nb_classes))
yname = "y_target"
else:
if viz_enabled:
# Initialize our array for grid visualization
grid_shape = (nb_classes, 2, img_rows, img_cols, nchannels)
grid_viz_data = np.zeros(grid_shape, dtype='f')
adv_inputs = x_test[idxs]
else:
adv_inputs = x_test[:source_samples]
adv_ys = None
yname = "y"
if targeted:
cw_params_batch_size = source_samples * nb_classes
else:
cw_params_batch_size = source_samples
cw_params = {'binary_search_steps': 1,
yname: adv_ys,
'max_iterations': attack_iterations,
'learning_rate': CW_LEARNING_RATE,
'batch_size': cw_params_batch_size,
'initial_const': 10}
adv = cw.generate_np(adv_inputs,
**cw_params)
eval_params = {'batch_size': np.minimum(nb_classes, source_samples)}
if targeted:
adv_accuracy = model_eval(
sess, x, y, preds, adv, adv_ys, args=eval_params)
else:
if viz_enabled:
err = model_eval(sess, x, y, preds, adv, y_test[idxs], args=eval_params)
adv_accuracy = 1 - err
else:
err = model_eval(sess, x, y, preds, adv, y_test[:source_samples],
args=eval_params)
adv_accuracy = 1 - err
if viz_enabled:
for j in range(nb_classes):
if targeted:
for i in range(nb_classes):
grid_viz_data[i, j] = adv[i * nb_classes + j]
else:
grid_viz_data[j, 0] = adv_inputs[j]
grid_viz_data[j, 1] = adv[j]
print(grid_viz_data.shape)
print('--------------------------------------')
# Compute the number of adversarial examples that were successfully found
print('Avg. rate of successful adv. examples {0:.4f}'.format(adv_accuracy))
report.clean_train_adv_eval = 1. - adv_accuracy
# Compute the average distortion introduced by the algorithm
percent_perturbed = np.mean(np.sum((adv - adv_inputs)**2,
axis=(1, 2, 3))**.5)
print('Avg. L_2 norm of perturbations {0:.4f}'.format(percent_perturbed))
# Close TF session
sess.close()
# Finally, block & display a grid of all the adversarial examples
if viz_enabled:
import matplotlib.pyplot as plt
_ = grid_visual(grid_viz_data)
return report